Building an Operational Brain: Skills, Subagents, Model Tiers, and the Token Economy of Delegated AI Work

Spanish title: Construir un cerebro operativo: skills, subagentes, niveles de modelo y la economía de tokens del trabajo delegado con IA Abstract: Most people who adopt a language-model assistant at work never get past prompting. The distance between a good prompt and a delegated process is not talent; it is structure. This paper describes that structure, which we call an operational brain: a durable configuration of skills, subagents, connectors, and model tiers that turns a general-purpose assistant into a coordinated system for day-to-day work. The contribution is architectural rather than empirical. We define a three-tier model of language engines — large (LLM), small (SLM), and domain-specific (DSLM) — and show that task selection and model selection are the same decision, not two. We identify a grounding gap as the root cause of everyday failure, drawing on common-ground theory, and derive a taxonomy of three recurring failure modes with a specific countermeasure for each. We then specify the six-layer architecture, the atomization contract that keeps delegation verifiable, and the three levers of token economy: context, cache, and model routing. Cached input is documented at roughly one tenth of the standard input price, which makes configuration files — not chat history — the economically rational place to store operational knowledge. Finally, we describe a three-layer self-learning loop that converts session failures into durable rules. The framework is offered as a reference pattern for practitioners; its quantitative claims are labelled as operational observations from a single deployment and are not presented as controlled results. Keywords: large language models; small language models; domain-specific language models; multi-agent delegation; token efficiency; prompt caching; knowledge management; human oversight. --- Languages: English (eng) and Spanish (spa) — the two manuscripts are line-for-line mirrors. Files: Markdown sources (containing the Mermaid source of every diagram), PDFs, and the rendered diagrams as PNG (figure-01…figure-09, one set per language). Conceptual/theoretical article; no dataset or software is required to apply the proposed framework. Not peer reviewed.